Project2Task proposes a planning layer for autonomous research agents that turns a broad research agenda into bounded, dependency-aware tasks.
The paper argues that a research project is not just a larger single task. It contains related objectives, parallel alternatives, and sequences that depend on earlier results. Existing systems may treat the project as one oversized job or produce vague overlapping task lists that still require manual coordination.
Project2Task represents candidate contributions as innovation atoms and organizes them in a directed lineage graph. It then selects a portfolio decomposition, repairs overlaps and missing fields, and emits task contracts with objectives, inputs, expected artifacts, and evaluation requirements.
The value is coordination. Research agents that can search literature, generate code, run experiments, and draft papers still need a plan that keeps work bounded and traceable. The method does not prove autonomous research is solved, but it addresses a practical bottleneck between a high-level idea and executable work.